Also to aid debugging, add sysfs support for iseries_veth's port structures.
Signed-off-by: Michael Ellerman <michael@ellerman.id.au>
Signed-off-by: Jeff Garzik <jgarzik@pobox.com>
To aid in field debugging, add sysfs support for iseries_veth's connection
structures. At the moment this is all read-only, however we could think about
adding write support for some attributes in future.
Signed-off-by: Michael Ellerman <michael@ellerman.id.au>
Signed-off-by: Jeff Garzik <jgarzik@pobox.com>
There's a number of problems with the way iseries_veth counts TX errors.
Firstly it counts conditions which aren't really errors as TX errors. This
includes if we don't have a connection struct for the other LPAR, or if the
other LPAR is currently down (or just doesn't want to talk to us). Neither
of these should count as TX errors.
Secondly, it counts one TX error for each LPAR that fails to accept the packet.
This can lead to TX error counts higher than the total number of packets sent
through the interface. This is confusing for users.
This patch fixes that behaviour. The non-error conditions are no longer
counted, and we introduce a new and I think saner meaning to the TX counts.
If a packet is successfully transmitted to any LPAR then it is transmitted
and tx_packets is incremented by 1.
If there is an error transmitting a packet to any LPAR then that is counted
as one error, ie. tx_errors is incremented by 1.
Signed-off-by: Michael Ellerman <michael@ellerman.id.au>
Signed-off-by: Jeff Garzik <jgarzik@pobox.com>
The iseries_veth driver often has multiple netdevices sending packets over
a single connection to another LPAR. If the bandwidth to the other LPAR is
exceeded, all the netdevices must have their queues stopped.
The current code achieves this by queueing one incoming skb on the
per-netdevice port structure. When the connection is able to send more packets
we iterate through the port structs and flush any packet that is queued,
as well as restarting the associated netdevice's queue.
This arrangement makes less sense now that we have per-connection TX timers,
rather than the per-netdevice generic TX timer.
The new code simply detects when one of the connections is full, and stops
the queue of all associated netdevices. Then when a packet is acked on that
connection (ie. there is space again) all the queues are woken up.
Signed-off-by: Michael Ellerman <michael@ellerman.id.au>
Signed-off-by: Jeff Garzik <jgarzik@pobox.com>
Currently the iseries_veth driver contravenes the specification in
Documentation/networking/driver.txt, in that if packets are not acked by
the other LPAR they will sit around forever.
This patch adds a per-connection timer which fires if we've had no acks for
five seconds. This is superior to the generic TX timer because it catches
the case of a small number of packets being sent and never acked.
This fixes a bug we were seeing on real systems, where some IPv6 neighbour
discovery packets would not be acked and then prevent the module from being
removed, due to skbs lying around.
Signed-off-by: Michael Ellerman <michael@ellerman.id.au>
Signed-off-by: Jeff Garzik <jgarzik@pobox.com>
The iseries_veth driver uses the generic TX timeout watchdog, however a better
solution is in the works, so remove this code.
Signed-off-by: Michael Ellerman <michael@ellerman.id.au>
Signed-off-by: Jeff Garzik <jgarzik@pobox.com>
The iseries_veth driver can attach to multiple vlans, which correspond to
multiple net devices. However there is only 1 connection between each LPAR,
so the connection structure may be shared by multiple net devices.
This makes module removal messy, because we can't deallocate the connections
until we know there are no net devices still using them. The solution is to
use ref counts on the connections, so we can delete them (actually stop) as
soon as the ref count hits zero.
This patch fixes (part of) a bug we were seeing with IPv6 sending probes to
a dead LPAR, which would then hang us forever due to leftover skbs.
Signed-off-by: Michael Ellerman <michael@ellerman.id.au>
Signed-off-by: Jeff Garzik <jgarzik@pobox.com>
This patch makes veth_init_connection() and veth_destroy_connection()
symmetrical in that they allocate/deallocate the same data.
Currently if there's an error while initialising connections (ie. ENOMEM)
we call veth_module_cleanup(), however this will oops because we call
driver_unregister() before we've called driver_register(). I've never seen
this actually happen though.
So instead we explicitly call veth_destroy_connection() for each connection,
any that have been set up will be deallocated.
We also fix a potential leak if vio_register_driver() fails.
Signed-off-by: Michael Ellerman <michael@ellerman.id.au>
Signed-off-by: Jeff Garzik <jgarzik@pobox.com>
The iseries_veth driver unconditionally calls dma_unmap_single() even
when the corresponding dma_map_single() may have failed.
Rework the code a bit to keep the return value from dma_unmap_single()
around, and then check if it's a dma_mapping_error() before we do
the dma_unmap_single().
Signed-off-by: Michael Ellerman <michael@ellerman.id.au>
Signed-off-by: Jeff Garzik <jgarzik@pobox.com>
The iseries_veth driver uses atomic ops to manipulate the in_use field of
one of its per-connection structures. However all references to the
flag occur while the connection's lock is held, so the atomic ops aren't
necessary.
Signed-off-by: Michael Ellerman <michael@ellerman.id.au>
Signed-off-by: Jeff Garzik <jgarzik@pobox.com>
The iseries_veth driver keeps a stack of messages for each connection
and a lock to protect the stack. However there is also a per-connection lock
which makes the message stack lock redundant.
Remove the message stack lock and document the fact that callers of the
stack-manipulation functions must hold the connection's lock.
Signed-off-by: Michael Ellerman <michael@ellerman.id.au>
Signed-off-by: Jeff Garzik <jgarzik@pobox.com>
Due to a logic bug, once promiscuous mode is enabled in the iseries_veth
driver it is never disabled.
The driver keeps two flags, promiscuous and all_mcast which have exactly the
same effect. This is because we only ever receive packets destined for us,
or multicast packets. So consolidate them into one promiscuous flag for
simplicity.
Signed-off-by: Michael Ellerman <michael@ellerman.id.au>
Signed-off-by: Jeff Garzik <jgarzik@pobox.com>
The iseries_veth driver contains a state machine which is used to manage
how connections are setup and neogotiated between LPARs.
If one side of a connection resets for some reason, the two LPARs can get
stuck in a race to re-setup the connection. This can lead to the connection
being declared dead by one or both ends. In practice the connection is
declared dead by one or both ends approximately 8/10 times a connection is
reset, although it is rare for connections to be reset.
(an example here: http://michael.ellerman.id.au/files/misc/veth-trace.html)
The core of the problem is that the end that resets the connection doesn't
wait for the other end to become aware of the reset. So the resetting end
starts setting the connection back up, and then receives a reset from the
other end (which is the response to the initial reset). And so on.
We're severely limited in what we can do to fix this. The protocol between
LPARs is essentially fixed, as we have to interoperate with both OS/400
and old Linux drivers. Which also means we need a fix that only changes the
code on one end.
The only fix I've found given that, is to just blindly sleep for a bit when
resetting the connection, in the hope that the other end will get itself
sorted. Needless to say I'd love it if someone has a better idea.
This does work, I've so far been unable to get it to break, whereas without
the fix a reset of one end will lead to a dead connection ~8/10 times.
Signed-off-by: Michael Ellerman <michael@ellerman.id.au>
Signed-off-by: Jeff Garzik <jgarzik@pobox.com>
The iseries_veth driver has a timer which we use to send acks. When the
connection is reset or stopped we need to delete the timer.
Currently we only call del_timer() when resetting a connection, which means
the timer might run again while the connection is being re-setup. As it turns
out that's ok, because the flags the timer consults have been reset.
It's cleaner though to call del_timer_sync() once we've dropped the lock,
although the timer may still run between us dropping the lock and calling
del_timer_sync(), but as above that's ok.
Signed-off-by: Michael Ellerman <michael@ellerman.id.au>
Signed-off-by: Jeff Garzik <jgarzik@pobox.com>
Currently the iseries_veth driver prints the file name and line number in its
error messages. This isn't very useful for most users, so just print
"iseries_veth: message" instead.
- convert uses of veth_printk() to veth_debug()/veth_error()/veth_info()
- make terminology consistent, ie. always refer to LPAR not lpar
- be consistent about printing return codes as %d not %x
- make format strings fit in 80 columns
Signed-off-by: Michael Ellerman <michael@ellerman.id.au>
Signed-off-by: Jeff Garzik <jgarzik@pobox.com>
My patch from a few weeks back (now in mainline), called "Cleanup skbs to
prevent unregister_netdevice() hanging", can cause our TX timeout code to
fire on machines with lots of VLANs (because it takes > 2 seconds between
when we stop the queues and when we're finished stopping the connections).
When that happens the TX timeout code freaks out and does a WARN_ON()
because as far as it's concerned there shouldn't be a TX timeout happening,
which is fair enough.
I have a "proper" fix for this, which is to a) do refcounting on
connections and b) implement a proper ack timer so we don't keep unacked
skbs lying around for ever. But for 2.6.12 I propose just supressing the
WARN_ON(). Users will still see the "NETDEV WATCHDOG" warning, but that's
not nearly as bad as a WARN_ON() which users interpret as an Oops.
Signed-off-by: Michael Ellerman <michael@ellerman.id.au>
Signed-off-by: Andrew Morton <akpm@osdl.org>
Signed-off-by: Linus Torvalds <torvalds@osdl.org>
Hi Andrew, Jeff,
The iseries_veth driver is badly behaved in that it will keep TX packets
hanging around forever if they're not ACK'ed and the queue never fills up.
This causes the unregister_netdevice code to wait forever when we try to take
the device down, because there's still skbs around with references to our
struct net_device.
There's already code to cleanup any un-ACK'ed packets in veth_stop_connection()
but it's being called after we unregister the net_device, which is too late.
The fix is to rearrange the module exit function so that we cleanup any
outstanding skbs and then unregister the driver.
Signed-off-by: Michael Ellerman <michael@ellerman.id.au>
Hi Andrew, Jeff,
Under some strange circumstances the iseries_veth driver can leak skbs.
Fix is simply to call dev_kfree_skb() in the right place.
Fix up the comment as well.
Signed-off-by: Michael Ellerman <michael@ellerman.id.au>
Hi Andrew, Jeff,
The iseries_veth driver doesn't set dev->trans_start in it's TX path. This
will cause the net device watchdog timer to fire earlier than we want it to,
which causes the driver to needlessly reset its connections to other LPARs.
Signed-off-by: Michael Ellerman <michael@ellerman.id.au>
Hi Andrew, Jeff,
The iseries_veth driver has a logic bug which means it will erroneously
send packets to LPARs for which we don't have a connection.
This usually isn't a big problem because the Hypervisor call fails
gracefully and we return, but if packets are TX'ed during the negotiation
of the connection bad things might happen.
Regardless, the right thing is to bail early if we know there's no
connection.
Signed-off-by: Michael Ellerman <michael@ellerman.id.au>
Initial git repository build. I'm not bothering with the full history,
even though we have it. We can create a separate "historical" git
archive of that later if we want to, and in the meantime it's about
3.2GB when imported into git - space that would just make the early
git days unnecessarily complicated, when we don't have a lot of good
infrastructure for it.
Let it rip!
Michael Ellerman